KR102645536B1 - Animation processing methods and devices, computer storage media, and electronic devices - Google Patents

Abstract

This disclosure provides an animation processing method and apparatus, and relates to the field of artificial intelligence. The method includes obtaining terrain features in a graphical user interface at the current moment, and obtaining state information and task information corresponding to a virtual character in an animation segment at the current moment; Input terrain features, state information, and task information into the animation processing model, and use the animation processing model to perform feature extraction on the terrain features, state information, and task information to obtain joint action information corresponding to the virtual character at the next moment. steps; And determining the joint torque according to the joint action information, performing rendering based on the joint torque to obtain gesture adjustment information corresponding to the virtual character at the current moment, and processing the animation segment according to the gesture adjustment information. .

Description

Animation processing methods and devices, computer storage media, and electronic devices

This disclosure claims priority to Chinese Patent Application No. 202010043321.5 filed on January 15, 2020, the content of which is incorporated by reference in its entirety.

The present disclosure relates to the field of artificial intelligence technology, and particularly to animation processing methods, animation processing devices, computer storage media, and electronic devices.

With the continued development of artificial intelligence, artificial intelligence technology is beginning to be applied to more and more fields, such as the medical field, financial field, and graphic design field. For example, game design evolves gradually from original 2D game design to current 3D game design.

Currently, in game production, a plurality of animation segments are generally designed by an animator, and the plurality of animation segments are mixed and converted using a game engine to finally achieve an effect in the game. do. Animation is an expression of character behavior, and a complete animation segment is one in which the actions of a character object over a certain period of time are recorded and played back. However, animations produced by animators have less natural and lifelike playback effects than animations rendered in real time in a physical engine, cannot interact with the player, and cannot, for example, implement variable target tasks. and cannot adapt to dynamic terrain.

The information disclosed in the foregoing background is merely to enhance understanding of the background of the present disclosure and may therefore include information that does not constitute related art known to those skilled in the art.

Embodiments of the present disclosure provide an animation processing method, an animation processing apparatus, a computer storage medium, and an electronic device.

Embodiments of the present disclosure provide an animation processing method applicable to electronic devices. The animation processing method includes obtaining a terrain feature from a graphical user interface at a current moment, and obtaining state information and task information corresponding to a virtual character from an animation segment at the current moment. ; Input the terrain features, the state information, and the task information into an animation processing model, perform feature extraction on the terrain features, the state information, and the task information using the animation processing model, and perform feature extraction on the terrain features, the state information, and the task information, and Obtaining joint action information corresponding to the character; determining joint torque according to the joint action information; and obtaining gesture adjustment information corresponding to the virtual character at the current moment based on the joint torque, and processing the animation segment according to the gesture adjustment information.

An embodiment of the present disclosure provides an animation processing device, wherein the animation processing device acquires terrain features in a graphical user interface at a current moment and acquires state information and task information corresponding to a virtual character in an animation segment at the current moment. an information acquisition module configured to; Input the terrain features, the state information, and the task information into an animation processing model, perform feature extraction on the terrain features, the state information, and the task information using the animation processing model, and perform feature extraction on the terrain features, the state information, and the task information, and a model processing module configured to acquire joint action information corresponding to the character; and a gesture configured to determine a joint torque according to the joint action information, obtain gesture adjustment information corresponding to the virtual character at the current moment based on the joint torque, and process the animation segment according to the gesture adjustment information. Includes coordination module.

An embodiment of the present disclosure provides a computer-readable storage medium that stores a computer program, and when the computer program is executed by a processor, it implements the animation processing method according to the above-described embodiment.

Embodiments of the present disclosure provide an electronic device, the electronic device comprising: one or more processors; and a storage device configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the animation processing method according to the above-described embodiment.

In the technical solutions provided in the embodiments of the present disclosure, the terrain features in the graphical user interface at the current moment and the state information and task information corresponding to the virtual character in the animation segment at the current moment are first obtained; Then, use the animation processing model to perform feature extraction on the terrain features, state information, and task information to obtain joint action information corresponding to the virtual character at the next moment; Finally, the joint torque is determined according to the joint action information, gesture adjustment information corresponding to the virtual character at the current moment is obtained based on the joint torque, and the animation segment is processed according to the gesture adjustment information. In this way, not only can animation segments be simulated, but also the actions and gestures of the virtual character can be adjusted according to different terrain features and task information. On the one hand, the fidelity of animation is improved; Meanwhile, interaction between the user and the virtual character is realized, and the self-adaptivity of the virtual character is improved.

It should be understood that the above general description and the following detailed description are for illustrative and explanatory purposes only and are not intended to limit the disclosure.

The drawings attached herein are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure. Obviously, the attached drawings described below are only some embodiments of the present disclosure, and those skilled in the art will be able to further obtain other attached drawings according to the attached drawings without creative efforts. In the attached drawing:
1 is a schematic diagram of an example system architecture to which the technical solutions of embodiments of the present disclosure can be applied.
Figure 2 schematically shows the configuration structure of a virtual character in conventional skinned animation.
Figure 3 schematically shows a schematic flowchart of an animation processing method according to an embodiment of the present disclosure.
Figure 4 schematically shows a schematic diagram of a scene obtained after a game scene and a real scene are integrated according to an embodiment of the present disclosure.
Figure 5 schematically illustrates a schematic diagram of an interface with densely-notched terrain according to an embodiment of the present disclosure.
Figure 6 schematically shows a schematic diagram of an interface of a hybrid-obstacle terrain according to an embodiment of the present disclosure.
Figure 7 schematically shows action information of a first frame of a walking action of a human-shaped character according to an embodiment of the present disclosure.
Figure 8 schematically shows a schematic diagram of a terrain interface according to an embodiment of the present disclosure.
Figure 9 schematically shows a schematic structural diagram of an animation processing model according to an embodiment of the present disclosure.
Figure 10 schematically shows a schematic structural diagram of a first control network according to an embodiment of the present disclosure.
Figure 11 schematically shows a schematic structural diagram of a second control network according to an embodiment of the present disclosure.
Figure 12 schematically shows a schematic flowchart of reinforcement learning according to an embodiment of the present disclosure.
Figure 13 schematically shows an architecture diagram of an algorithmic framework of an animation processing model according to an embodiment of the present disclosure.
Figure 14 schematically shows an action sequence of a virtual character running on flat ground and controlled by an animation processing model, according to an embodiment of the present disclosure.
15A to 15E schematically illustrate an action sequence of a human-shaped virtual character running on dense notched terrain according to an embodiment of the present disclosure.
16A to 16L schematically show an action sequence of a human-shaped virtual character running on a hybrid obstacle terrain according to an embodiment of the present disclosure.
Figure 17 schematically shows a block diagram of an animation processing device according to an embodiment of the present disclosure.
Figure 18 shows a schematic structural diagram of a computer system of an electronic device adapted to implement an embodiment of the present disclosure.

An example implementation will now be described more thoroughly with reference to the accompanying drawings. However, the present disclosure may be implemented in various forms and is not limited to the embodiments described herein. Instead, implementations are provided so that this disclosure will be more thorough and complete and will fully convey the ideas of the example implementations to those skilled in the art.

Additionally, the described features, structures or characteristics may be combined in one or more embodiments in any suitable way. In the following description, numerous details are provided to obtain a thorough understanding of embodiments of the present disclosure. However, those skilled in the art should recognize that the technical solutions of the present disclosure may be implemented without one or more specific details, or that other methods, units, devices or steps may be used. In other instances, well-known methods, devices, implementations or operations are not shown or described in detail so as not to obscure aspects of the disclosure.

The block diagrams shown in the attached drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, the functional entity may be implemented in software form, or as one or more hardware modules or integrated circuits, or as a different network and/or processor device and/or microcontroller device.

The flowcharts shown in the accompanying drawings are merely illustrative illustrations and do not necessarily include all content and operations/steps, nor do they necessarily need to be performed in the order described. For example, some operations/steps may be further divided while some operations/steps may be combined or partially combined. Therefore, the actual execution order may change depending on the actual case.

1 is a schematic diagram of an example system architecture to which the technical solutions of embodiments of the present disclosure can be applied.

As shown in Figure 1, system architecture 100 may include a terminal device 101, a network 102, and a server 103. Network 102 is a medium used to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired communication links, wireless communication links, etc.

It should be understood that the number of terminal devices, number of networks, and number of servers in FIG. 1 are only examples. There may be any number of terminal devices, any number of networks, and any number of servers, depending on actual requirements. For example, server 103 may be a server cluster including a plurality of servers. The terminal device 101 may be a terminal device including a display screen, such as a laptop, portable computer, or desktop computer.

In an embodiment of the present disclosure, a game application is carried in the terminal device 101, and the game application includes an animation segment. While running the gaming application, obstacles may be set for the virtual character using the relevant controls of the gaming application; Alternatively, an actual scene may be photographed using the photographing unit of the terminal device 101, and an obstacle for the virtual character may be set by integrating the actual scene into the game screen. Meanwhile, the user can set a task for the virtual character according to the scene of the animation segment, for example, have the virtual character move in a target direction or toward a target point. The terminal device 101 uses the network 102 to send terrain features in the graphical user interface at the current moment, and task information and status information corresponding to the virtual character in the animation segment at the current moment, to the server 103. The terrain features, task information, and status information are processed using the server 103 to obtain gesture adjustment information corresponding to the virtual character at the current moment. In this way, not only can animated segments be simulated, but virtual characters can also be self-adaptive and complete set tasks.

In some embodiments, feature extraction may be performed on terrain features, state information, and task information using an animation processing model to obtain joint action information corresponding to the virtual character at the next moment; Joint torque may be determined based on joint action information, and the joint torque is applied to the corresponding joint using a physical engine to perform rendering and gesture adjustment information corresponding to the virtual character at the current moment. can be obtained. The state information corresponding to the virtual character may be gesture information corresponding to the virtual character at an initial moment of the animation segment, or may be state information determined based on joint action information at a previous moment. An animation segment has a certain duration, gesture adjustment information corresponding to the virtual character at a plurality of moments can be obtained by repeating the above-described operation, and a target action sequence is at a plurality of moments. It can be determined according to the gesture coordination information, the target action sequence can form an animation segment, the animation segment is similar to the animation segment of the running game and has high fidelity, the difference is that the virtual character in the animation segment is set by the user. It is capable of self-adapting to the terrain and completing tasks set by the user, that is, the technical solution of embodiments of the present disclosure improves the interaction between the user and the virtual character. By improving the self-adaptability of virtual characters, the user experience can be further improved.

The animation processing method provided in the embodiment of the present disclosure may be performed on a server, and correspondingly, an animation processing device may be placed on the server. However, in another embodiment of the present disclosure, the animation processing method provided in the embodiment of the present disclosure may otherwise be performed by a terminal device.

Server 103 may be an independent physical server, or a server cluster including multiple physical servers or distributed systems, or may be a cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud It may be a cloud server that provides basic cloud computing services such as communications, middleware services, domain name services, security services, content delivery network (CDN), big data, and artificial intelligence platforms.

In related technologies in this field, 3D games can be taken as an example, and character animation in 3D games generally means skinned animation. Figure 2 shows the composition structure of a virtual character in a skinned animation. As shown in Figure 2, the virtual character of the skinned animation includes a bone, a skin, and an animation, and the bone is a movable skeleton composed of joints and a virtual subject that can also move, allowing the entire character to move. However, it is not rendered in game; A skin is a triangular mesh that surrounds a bone, and each vertex of the mesh is controlled by one or more bones. Animation is a change in the position or direction of each bone at a specific time point, and the three-dimensional space is a general It is expressed as a matrix. Typically, animators use 3D animation production software to design and produce a large number of animation segments in advance, and while playing the game, the program plays the animation segments needed for the scene at the appropriate time. In particular, if necessary, the program performs post-animation post-processing before rendering, for example, the exact positioning of the virtual character's hands and feet can be determined using inverse kinematics (IK) methods to coordinate the action with the real environment at that time. It is calculated accordingly. However, the effectiveness of post-processing is limited, and the quality of the animation generally depends almost entirely on the skill of the animator. When an animator produces animation directly, it means that the animation is actually played directly in the game. The physical laws of the real world are not simulated by a physical engine, so the character's actions are not natural or lifelike enough. There are currently several machine learning solutions in the industry for training physically animated AI. However, the learning effect is generally poor, and one model can only learn one action with a single performance.

Additionally, the main method of implementing animation in modern game production is to play animation segments produced by animators, which can practically only be applied to predefined enumerable scenes and have no self-adaptation ability to the environment. The character's self-adaptivity to the environment means that the character animation can present gestures matching the environment in an unknown environment. The 'unknown' here relates to the environment assumed during the pre-production process of the animation; The real world may have large or small changes while using an animated segment. Additionally, collisions can be recognized due to external interference, suggesting deviations and corrections in actions that have a scene with a strong sense of realism. At the very least, IK technology should realize self-adaptation to the environment and allow the character's extremities to be aligned with the environment or target in terms of position; And if the character's feedback about the environment must be vivid in order to calculate the appropriate speed of the character's actions and a smooth transition process, additional "physics" (i.e. simulation of rigid body dynamics) need to be introduced. . In general, the terrain is fixed, the action process of the character moving through the terrain is animated, and unnatural parts are appropriately modified. This procedure is essentially still intended to play animations, and the character's movement across the terrain is unnatural.

Embodiments of the present disclosure provide an animation processing method in consideration of existing problems in related technologies. This method is implemented based on artificial intelligence (AI). Artificial intelligence uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to produce optimal results. It is a system of theories, methods, techniques and applications to be acquired. In other words, AI is a comprehensive technology of computer science that seeks to understand the nature of intelligence and produce new types of intelligent machines that can respond similarly to human intelligence. AI is the study of design principles and implementation methods of various intelligent machines to enable machines to have perception, reasoning, and decision-making functions.

AI technology is a comprehensive study that covers a wide range of fields, including both hardware-level technology and software-level technology. Basic AI technologies typically include technologies such as sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, and mechatronics. AI software technology mainly includes computer vision technology, voice processing technology, natural language processing technology, and machine learning/deep learning.

Computer vision (CV) is a method of enabling a machine to "see", specifically allowing a computer to process a target into an image that is better suited for observation by the human eye or transmitted to a detection device. It is a science that studies how to implement machine vision such as target recognition, tracking, and measurement using cameras and computers that replace the human eye and perform additional graphic processing. CV is a scientific subject that studies related theories and technologies and attempts to build an AI system that can obtain information from images or multidimensional data. CV technologies typically include image processing, image recognition, image semantic understanding, image retrieval, optical character recognition (OCR), video processing, video semantic understanding, video content/action recognition, and 3D object reconstruction. , 3D technology, virtual reality, augmented reality, synchronous positioning, and map construction, and further includes biometric feature recognition technologies such as common face recognition and fingerprint recognition.

ML is a multi-field interdiscipline and concerns multiple academic fields such as probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. ML specializes in studying how computers simulate or implement human learning actions in order to continuously improve computer performance by acquiring new knowledge or skills and reconstructing existing knowledge structures. ML is the core of AI and is the basic method of making computers intelligent and is applied to various fields of AI. ML and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning from demonstrations.

With the research and development of artificial intelligence technology, artificial intelligence technology is used in various fields such as general smart home, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, unmanned aerial vehicles, robots, smart medicine, and smart customer service. It is being researched and applied. It is believed that with the advancement of technology, AI technology will be applied to more fields and play an increasingly important role.

The solution provided in the embodiment of the present disclosure includes AI image processing technology, and is described using the following embodiment.

Embodiments of the present disclosure first provide an animation processing method. Figure 3 schematically shows a flowchart of an animation processing method according to an embodiment of the present disclosure. The animation processing method may be performed on a server, and the server may be the server 103 shown in FIG. 1. The processing of game animation is used as an example. Referring to FIG. 3, the animation processing method includes at least steps S310 to S330.

In step S310: terrain features are obtained from the graphical user interface at the current moment, and state information and task information corresponding to the virtual character are obtained from the animation segment at the current moment.

In an embodiment of the present disclosure, in order to improve the fun of the game and enhance the interaction between the user and the virtual character in the game, the user artificially sets obstacles for the virtual character during the game, and creates a new device in the graphical user interface. The terrain can be set, for example, if the virtual character walks straight along a flat road in the original animation segment, the user can set roadblocks in the virtual character's moving path, roadblocks can be made of stones. , may be an obstacle such as a step or a pit; Alternatively, the user can set obstacles in the sky, such as eaves or flying birds, in the virtual character's movement path, and the virtual character must avoid these obstacles to continue moving forward. In order to clarify the technical solution of the embodiment of the present disclosure, the following description takes a roadblock as an example, and the roadblock may be an obstacle such as a notch, protuberance, or step in the ground.

In an embodiment of the present disclosure, the roadblock set by the user may be set using controls placed inside the game, or may be set according to the actual scene. In some embodiments, a roadblock setting button may be set in the game interaction interface. When the user triggers the roadblock setup button, a list may pop up, and the user selects the roadblock they want to set for the virtual character from the list. After the user's decision, the corresponding roadblock appears in the game interface. In an augmented reality game, the actual scene is filmed using a photography unit provided in the terminal device used by the user, and the game engine can integrate the real scene and the game scene. Figure 4 shows a schematic diagram of the scene obtained after the game scene and the real scene are integrated. As shown in Figure 4, there is a demon spirit (V) in the game scene, a plurality of steps (S) in the actual scene, and a plurality of electric scooters on the platform of the top step. (M) is placed. The evil spirit (V) can be placed on the step (S) of the real scene by integrating the game scene and the real scene.

In an embodiment of the present disclosure, setting obstacles in a game is generally setting roadblocks, and the terrain generated using a relatively large number of obstacles can be called densely-notched terrain and hybrid obstacle terrain ( hybrid-obstacle terrain). Figure 5 shows a schematic diagram of the interface of the dense notch topography. As shown in FIG. 5, the dense notch topography is designed so that there are a plurality of continuous notches on the ground (G), the notches have different widths, and there is a constant interval between the two notches. Figure 6 shows a schematic diagram of the interface of the hybrid obstacle terrain. As shown in FIG. 6, the hybrid obstacle terrain includes obstacles such as notches (C), steps (D), and protrusions (E) on the ground (G) of a certain length. The height and width of the obstacles are different, and there is a certain gap between the two obstacles.

In an embodiment of the present disclosure, in addition to setting a roadblock for the virtual character, a task for the virtual character may be additionally set while the virtual character is moving. For example, there is a soccer ball in front of the virtual character, the task can be set to “kick the soccer ball,” and the coordinate position of the soccer ball can be used as the target point to determine task information according to the target point. In order to drive the virtual character to move in a specific direction, a target speed direction is additionally set for the virtual character, and task information may be determined according to the target speed direction.

In embodiments of the present disclosure, the virtual character's gestures and actions are associated with each other in continuous time and space. Using a human-shaped virtual character taking a step as an example, if the human-shaped virtual character's right foot is lifted at the current moment, the right foot tends to land at the next moment. Therefore, the determination of the joint action information of the virtual character at the next moment should be based on processing the state information of the virtual character at the current moment, and the state information is used to describe the state of each joint of the virtual character, including the gestures of the joints, May include speed and phase. Therefore, to determine how to change the gestures of a virtual character at the current moment to avoid obstacles and complete a task, terrain features in the graphical user interface at the current moment and state information and task information corresponding to the virtual character at the current moment can be obtained, and this information can be processed to obtain corresponding gesture adjustment information.

In an embodiment of the present disclosure, when creating an animation segment, an animator may set the animation segment to different formats; When extracting the state information of a virtual character from an animation segment, first convert the format of the animation segment into a file in FBX format or BVH format using some software (e.g. MotionBuilder or 3ds Max), and then extract the state information. . In actual implementation, when the current moment is the initial moment of the animation segment, the state information is determined according to the gesture information of the virtual character at the initial moment of the animation segment, and the gesture information at the initial moment may be determined as the state information during implementation. There is; If the current moment is a moment other than the beginning of the animation segment, the state information is determined according to the joint action information corresponding to the virtual character at the previous moment, and during implementation, the joint action information corresponding to the virtual character at the previous moment is converted into state information. can be decided.

In an embodiment of the present disclosure, a human-shaped virtual character is used as an example, and the human-shaped virtual character has a total of 15 joints, respectively, root joint, chest, neck, right leg, left leg, right knee, These are the left knee, right ankle, left ankle, right shoulder, left shoulder, right elbow, left elbow, right hand and left hand, where the root joint usually refers to the pelvic position and is denoted as the root. In general, the bones and joints of human-shaped virtual characters have a parent-child hierarchical structure. For example, the shoulder is a parent joint, the elbow is a child joint of the shoulder, and the wrist is a child joint of the elbow. The position of the child joint is obtained by performing a corresponding translation from the position of the parent joint. Therefore, there is no need to record the position coordinates of the child joints, the position coordinates of the top root joint are known, and when the translation is performed according to the size of the bone set by the animator during animation design, the position coordinates of the child joints are obtained. can do. In the case of actions, by recording the gesture information of the character's joints in the animation segment and knowing the position and rotation of each necessary joint, the current action of the virtual character can be configured. In addition to the position and rotation of the root joint, the corresponding rotations of other joints are recorded to constitute the current overall gesture of the virtual character. Figure 7 shows action information of the first frame of a walking action of a human-shaped character. As shown in FIG. 7, the action information of the first frame is divided into three lines. The first number 0.0333333 in the first line represents the duration of the first frame in seconds, and the next three numbers (001389296, 0.8033880000000001, 0.0036694320000000002) represent the root joint of the first frame in three-dimensional space. coordinates; The four numbers in the second line (0.5306733251792894, -0.5324986777087051, -0.4638864011202557, -0.46865807049205305) are rotation information of the root joint of the first frame; The four numbers of the third row (0.7517762842400346, 0.0012912309982618, -0.0033 740637622359164, 0.6594083839744481) are the rotation of the first child joint corresponding to the root joints, and the remaining rotation information of the remaining child joints It is omitted in FIG. 7. Rotation information is expressed as a unit quaternion. Unit quaternions can be used to express rotations in three-dimensional space and are equivalent to the commonly used three-dimensional orthogonal matrices and Euler angles, but avoid the gimbal lock problem in Euler angle representation. If the Cartesian coordinates of a point in three-dimensional space are (x, y, z), then that point is a pure quaternion (which is similar to a pure imaginary number, that is, a real number of 0). A quaternion with components) is expressed as xi+yj+zk. The geometric meaning of i, j or k can be understood as a kind of rotation, where rotation i represents a rotation from the positive It represents a rotation from the positive Z-axis direction toward the positive X-axis direction in the intersection plane of the Y-axis and the Indicates that -i, -j, and -k represent the reverse rotation of rotation i, rotation j, and rotation k, respectively.

In an embodiment of the present disclosure, the state information input to the animation processing model may be a 197-dimensional vector, where the included gesture may be 106 dimensional, the included speed may be 90 dimensional, and the included phase may be 1 dimensional. there is. In some embodiments, the gesture records position and rotation information of 15 joints of a human-shaped character, where the positions are expressed in three-dimensional coordinates and the rotation information is expressed as a unit quaternion with 15*7 = 105 total dimensions. . Additionally, the one-dimensional y-axis value of the virtual character's root joint coordinates at the current moment must be additionally recorded, which is used for alignment with the world coordinate system. Speed records the linear speed and angular speed of each joint, and each is expressed as a vector with a length of 3 corresponding to the x-axis speed, y-axis speed, and z-axis speed, so 15*(3 +3)= There are a total of 90 dimensions. Phase records the position of the current moment in the total time length of the animation segment, which is in one dimension overall.

In an embodiment of the present disclosure, the terrain feature may be a two-dimensional matrix, where each element of the matrix is the relative height difference between the terrain height of the corresponding point and the height of the current location of the virtual character, which is a preset height in front of the virtual character. Covers the height of the region within the range. The size of the matrix and the area of covering terrain can be adjusted according to the actual application scenario. For example, the size of the two-dimensional matrix is set to 100*100, and the area of the terrain it covers is set to 10m*10m. This is not limited to the embodiments of the present disclosure. Figure 8 shows a schematic diagram of the terrain interface. As shown in FIG. 8, the terrain is a rectangular area, the virtual character is located at the midpoint of the left sideline, and the arrow indicates the movement direction of the virtual character A. Since virtual character A only moves forward in the horizontal direction without turning, is parallel to obstacle B and is vertically as high as obstacle B, the terrain features are a 100*1 matrix covering terrain features within 10m in front of the virtual character. can be decided.

When obtaining terrain features from the graphical user interface of the current moment and status information and task information corresponding to the virtual character from the animation segment of the current moment, the terrain features, status information, and task information of the current moment transmitted by the terminal can be received. Alternatively, the server itself may determine the current moment's terrain features, status information, and task information according to the graphical user interface and the received configuration information.

In step S320: Input terrain features, state information, and task information into the animation processing model, and use the animation processing model to perform feature extraction on the terrain features, state information, and task information, to generate the virtual character at the next moment. Obtain corresponding joint action information.

In an embodiment of the present disclosure, after obtaining the terrain features, state information, and task information at the current moment, this information may be input into an animation processing model, and the animation processing model may be used to determine the terrain features, state information, and task information. By performing feature extraction, joint action information corresponding to the virtual character at the next moment can be obtained. Joint action information is information about the actions that each joint can perform at the next moment when the virtual character faces the terrain features and task features at the current moment, and the joint action information may be the rotation information of the joint, which is a four-dimensional It is expressed in length and includes rotation information of joints other than the root joint, which is (15-1)*4 = 56 dimensions in total. The gesture of the root joint can be simulated and acquired by the physics engine after the other joints perform translation and rotation under the torque effect. For example, a human-shaped virtual character walks forward on level ground, and the physical engine performs movement and rotation according to the torque determined by rotation information of joints such as the legs and knees, and then the rear static friction force (backward) applied to the feet Static friction force) is sequentially transmitted to the lower extremities, knees, thighs, and root joints, and the root joint is pushed forward under the influence of the force. Accordingly, the action information of the root joint may be omitted from the joint action information.

In an embodiment of the present disclosure, feature extraction is performed on the acquired terrain features, state information, and task information using a reinforcement learning-based animation processing model to obtain joint action information corresponding to the virtual character at the next moment. You can. Figure 9 shows a schematic structural diagram of an animation processing model. As shown in Figure 9, the animation processing model 900 includes a first control network 901 and a second control network 902. The first control network 901 may be a high-level controller (HLC) configured to guide the actions of key joints of the virtual character, where the key joints include terrain features, the state of the virtual character, and the like. These are some joints that correspond to information and task information. For example, when a human-shaped virtual character is running, the action of the legs mainly changes, and the thighs move the legs and feet, so the main joint becomes the thigh joint. Similarly, when a human-shaped virtual character is throwing, the main joint is the elbow, as the action of the arm and hand changes and the elbow moves by moving the wrist and hand. The second control network 902 may be a low-level controller (LLC) configured to output joint action information corresponding to all joints. Adaptation to complex animation scenes and complex tasks can be better implemented by configuring the first control network and the second control network respectively. Additionally, the first control network is mainly configured to guide specific actions, and the second control network is mainly configured to control movement of the character. A plurality of first control networks for different specific actions may be connected to a trained second control network. For example, the trained second control network may output joint action information of the foot action of the virtual character according to the target state information of the foot, and the action corresponding to the target state of the foot may be an action of the virtual character kicking a ball with the foot or a virtual character kicking a ball with the foot. This could be an action where the character jumps. In this case, the same second control network may be connected to the first control network that guides the virtual character's kicking of the ball, and may also be connected to the first control network that guides the virtual character's jumping. Animation segments are processed using an animation processing model that includes a first control network and a second control network, which can improve the effect of the action and the fidelity of the action, and can adapt to various terrains, providing self-adaptation to the environment. improve

In some embodiments, the first control network 901 performs feature extraction on terrain features and state information and task information corresponding to the virtual character at the current moment to obtain target state information corresponding to key joints; Then, the target status information is determined as target task information, and the status information and target task information are input to the second control network 902. Feature extraction is performed on the state information and target task information corresponding to the virtual character using the second control network 902 to obtain joint action information corresponding to all joints of the virtual character. An example would be a case where a human-shaped virtual character overcomes an obstacle. In order for the human-shaped virtual character to successfully overcome obstacles of different heights, the human-shaped virtual character lifts its legs at different angles and takes steps at different sizes. The first control network 901 may output the rotation of the two thigh joints and the speed direction of the root joint on the plane of the character according to the terrain characteristics, task information, and state information, and the rotation and direction of the speed of the root joint on the plane of the character. The speed direction of the root joint is target state information corresponding to the main joint, and its output is used as a target task of the second control network 902 to guide the human-shaped virtual character to lift the leg. Correspondingly, the output of the first control network 901 may be a 10-dimensional vector, i.e., a unit quaternion measuring the rotation of both thighs and a unit vector of length 2. Of course, the target state information may be additional rotation of the two hand joints, rotation of the two shoulder joints, etc., in addition to the rotation of the two thigh joints and the speed direction of the root joint on the character's plane. Target status information varies depending on obstacle type and task information.

Additionally, Figure 10 shows a schematic structural diagram of the first control network. As shown in Figure 10, the first control network 901 includes a convolution unit 1001, a first fully connected layer 1002, a second fully connected layer 1003, and It includes a third fully connected layer 1004, where the convolution unit 1001 may include a plurality of convolution layers of different sizes. As shown in the figure, the size of the first convolution layer set is 8*8, the sizes of the second and third convolution layer sets are both 4*4, and the first fully connected layer (1002 ), the sizes of the second fully connected layer 1003 and the third fully connected layer 1004 are different from each other, where the first fully connected layer 1002, the second fully connected layer 1003 and the third fully connected layer ( The numbers of nerve cells included in 1004) are 64, 1024, and 512, respectively. After the terrain feature T, task information g _H and state information s _H are input to the first control network, feature extraction for the terrain feature T is first performed using the convolution unit 1001 to obtain first feature information; ; Then, using the first fully connected layer 1002, feature combination is performed on the first feature information to obtain second feature information; Next, using the second fully connected layer 1003, feature combination is performed on the second feature information, state information s _H , and task information g _H to obtain third feature information; Using the third fully connected layer, feature combination is finally performed on the third feature information to obtain target state information a _H.

Figure 11 shows a schematic structural diagram of the second control network. As shown in Figure 11, the second control network 902 includes a fourth fully connected layer 1101 and a fifth fully connected layer 1102, and the fourth fully connected layer 1101 and the fifth fully connected layer The sizes of the tiers are different. In some embodiments, the fourth fully connected layer 1101 may include 1024 nerve cells, and the fifth fully connected layer 1102 may include 512 nerve cells. After the first control network 901 outputs the target state information a _H , the target state information a _H may be regarded as the target task information g _L of the second control network 902, and the first control network 901 together with the state information s _L 2 are simultaneously input to the control network 902, and feature combination is performed on the state information s _L and target task information g _L using the fourth fully connected layer 1101 to obtain the fourth feature information, and then the fifth Joint action information a _L is obtained by performing feature combination on the fourth feature information using the fully connected layer 1102.

In an embodiment of the present disclosure, the first control network 901 guides the action of the main joints of the virtual character, that is, a specific action, and the second control network 902 outputs joint action information of all joints of the virtual character to , forms a continuous action, that is, controls the movement of the character, and accordingly, the call period of the first control network 901 and the second control network 902 is different. That is, the first control network 901 needs to be called only when the character's action or the state of the main joint changes; And as long as the virtual character moves, each joint corresponds to the corresponding joint action information, so the second control network 902 needs to be continuously called. An example would be a virtual character crossing a roadblock. While the first control network 901 must be called only when the virtual character takes a step, the second control network 902 is continuously called to control the virtual character to take continuous actions. By setting different call cycles for the first control network 901 and the second control network 902, time and resources can be saved, improving the processing efficiency of the animation processing model, and thus the efficiency of action generation. can be improved. In an embodiment of the present disclosure, the call frequency of the first control network 901 for the PD controller is 2 Hz, the call frequency of the second control network 902 is 30 Hz, and the physical simulation frequency is 3000 Hz. In actual use, it is determined whether the first control network 901 should be called according to the terrain features, task information, and status information at the current moment. However, the second control network 902 needs to be continuously called in order to predict the joint action information of the virtual character at the next moment. If the first control network 901 is not called, the input of the second control network 902 is not changed.

In an embodiment of the present disclosure, before performing feature extraction on terrain features, state information, and task information using the animation processing model, a to-be-trained animation processing model is used to obtain a stable animation processing model. There is a need to train. While training an animation processing model, a commonly adopted method is to input terrain features into the model. However, this method is only moderately effective, training is prone to failure, and the character's actions are a bit stiff, and as a result, this method only adapts to relatively simple terrain. Therefore, in an embodiment of the present disclosure, in order to strengthen the sensitivity of the animation processing model to terrain and action and migrate to a more complex terrain, hierarchical reinforcement learning is applied by splitting the terrain features input to the model. do. Reinforcement learning is a field of machine learning and emphasizes how to take actions based on the environment to maximize expected effects. The movement control problem is a standard benchmark in reinforcement learning, and deep reinforcement learning methods have been proven to be adaptable to several tasks, including manipulation and movement.

In an embodiment of the present disclosure, reinforcement learning includes a plurality of basic concepts, which are environment, intelligent agent, state, action, reward, value function, and policy, respectively. The environment is an external system, and an intelligent agent is located inside the system and can recognize the system and perform specific actions according to the recognized state. Intelligent agents are systems embedded in the environment and can perform tasks that change their state. The state is information about the state of the current environment at the moment. An action is an action performed by a subject. Reward is a scalar that represents the environment's reward for the current action or state. Reward defines the immediate benefit, while the value function defines the long-term benefit, which can be considered as an accumulated reward and is usually denoted by V. A policy is a mapping from the current state of the environment to an action, usually denoted by π, the input state, and the model outputs the action that should be performed in the state. Figure 12 shows a schematic flowchart of reinforcement learning. As shown in Figure 12, at moment t, the current state S _t is input to the intelligent agent, and according to the current policy, the intelligent agent can output the action A _t ; The action A _t is performed to interact with the environment, and according to the completion condition of the target, the environment feeds back the reward R _t and the state S _t+ 1 of the intelligent agent at the next moment t+1; The intelligent agent adjusts the policy according to the reward and outputs the action A _t+1 at the next moment; Then, you can repeat this process, continuously adjust the policy, and finally obtain a policy π that can complete the target through training.

In an embodiment of the present disclosure, the animation processing model is trained based on the AC framework. The AC framework is a framework that integrates a value function estimation algorithm and a policy search algorithm, and includes two networks: an actor network and a critic network. The actor network is configured to train the current policy and output an action, and the critic network is configured to learn the value function and output the current state value V(s), which evaluates the quality of the state. It is used to Figure 13 shows an architecture diagram of the algorithmic framework of the animation processing model. As shown in Figure 13, the framework includes an actor network 1301, a critic network 1302, and an environment 1303. The actor network 1301 outputs an action according to the current state and policy, the environment 1303 provides feedback in the form of compensation according to the action output by the actor network 1301, and the critic network 1302 determines whether the action was performed. Afterwards, an evaluation is performed according to the generated state and the reward fed back by the environment 1303, the current state value is determined, and the current state value is fed back to the actor network 1301 so that the actor network 1301 adjusts the policy. do. This process is repeated and training continues until the animation processing model is stable. A learning standard of the current state value output by the critical network 1302 is obtained by calculating a series of rewards fed back by the environment 1303 using a temporal difference method, And it is used to guide the learning of the critical network. In some embodiments, path simulation is used as an example, and rewards R ₁ to R _i corresponding to nodes on the path can be obtained, where i is the number of nodes on the path; If you want to find the state value V(S _t ) corresponding to node t on the path, t is a value in the range from 1 to i, and the value of S _t V(S _t ) is the obtained reward R and the state value of the subsequent state. is updated according to the estimated value of . After the update is performed several times, a stable value function is obtained; After sampling is performed on a path, the value function can be updated multiple times, and the evaluation algorithm used is It can be, and α is the coefficient. The temporal difference method is the central idea of reinforcement learning. Similar to the Monte Carlo method, the time-difference method combines the Monte Carlo method's sampling method (i.e., performing an experiment) and the dynamic programming method's bootstrapping (using the value function of the subsequent state to determine the current state). It is combined with an estimate of the value function) and can be known directly from experience without complete knowledge of the environment. Similar to the dynamic programming method, the temporal difference method can improve existing estimation results without waiting for the entire event to complete, thereby increasing learning efficiency.

In an embodiment of the present disclosure, during training of the model, the physics engine includes two characters, a kinematics character and a physical character, respectively. Kinematic characters have no physical properties and are only used to perform actions in action segments designed by animation designers, and are only needed to ensure that the joints of a kinematic character perform reference actions in animation segments using kinematic methods. Physical characters can be learned using kinematic characters as standards and templates, have physical properties, and can be controlled by torque, and physical properties can be torque, speed, gravity, collision effects, etc. there is. Meanwhile, for physical characters with physical properties, the torque of each joint is calculated using the gesture output by the model, and action simulation is performed in the physical engine. The physics engine creates realistic effects by simulating the conditions of the environment after each action is performed. At each moment, calculating the reward means measuring the difference between the current gestures, speeds, and angular velocities of the two characters. The smaller the difference, the greater the reward. The final reward is obtained by performing weighted summation on the plurality of reward components, and the weights can be adjusted as needed. The environment provides rewards based on the quality of the gesture simulation, stimulating the character to maintain gestures that are consistent with those of the reference action. The closer the two gestures are, the larger the reward; otherwise, the smaller the reward.

In an embodiment of the present disclosure, the compensation is determined according to Equation (1), and Equation (1) is as follows.

is the compensation value of the simulation at instant t, is the reward value for completing the target task at instant t, and is the weight represents the rate at which the action is simulated, and the weight is the rate at which the task is completed, in engineering and can be set, am.

In order to ensure that the actions of the physical and kinematic characters are consistent, some criteria can be set for the physical and kinematic characters, and the formula Five Parts: Gesture Compensation , speed compensation , extremity joint compensation , root joint gesture compensation and centroid gesture compensation It includes similarity in kinematic aspects including. Gesture and speed are the gesture and speed of each joint. If the actions of the two characters must match, the gesture and speed must match, so gesture compensation and speed compensation can be set. The limb joints represent the hands and feet. To ensure that the physical character's limb joints are aligned with the kinematic character's limb joints, limb joint compensation is set for the limb joints. The root joint is the top joint of all joints. If the actions of the two characters must match, the root joints must also match, and root joint gesture compensation can be set accordingly. Additionally, in order for the physical character to walk more stably and not shake, the centroid of the physical character must match the center of the kinematic character, so centroid gesture compensation can be set. By setting the above-described compensation, it can be ensured that the actions of the physical character and the kinematic character match as much as possible. The weight corresponding to the reward is am. The upper right corner of a kinematic character item is marked with *. A gesture component is used as an example, where is the gesture of the jth joint of the kinematic character, is the gesture of the jth joint of the simulation character. Formula (1) can be converted to equation (2) as follows.

here, describes the similarity between gestures, and is expressed as the difference between the position and rotation of each joint and the target value, and; describes the similarity between velocities, and is expressed as the difference between the linear velocity of each joint and the target value. and; describes the similarity between limb joint gestures, and is expressed as the difference between the positions of the hand joints and the positions of the foot joints, and; describes the similarity between root joints, ego; describes the similarity between the central velocities, am.

describes the quality with which a character achieves a target, and generally measures the difference between actual conditions and the target of the character's movement. For example, if the target has direction of movement g _t , Can be calculated as the angle difference θ between the forward direction v _t and the target g _t on the ground, as shown in the following equation (3).

If the virtual character fails to learn the action of falling down, the current training path is completed and the reward value is 0.

In an embodiment of the present disclosure, the animation processing model to be trained includes a first control network to be trained and a second control network to be trained. Before training, a plurality of animation segment samples can be obtained, the animation segment samples include different terrain features and different task information corresponding to the virtual character, and corresponding to the different terrain features and different task information, the virtual character Characters have different gestures and actions. In an embodiment of the present disclosure, the first control network of the animation processing model outputs target state information corresponding to major joints according to terrain features, task information, and state information, and then the target state information is used as target task information to control the second It is input to the network and processed by the second control network to output joint action information, which can be used to process complex tasks. If the first control network and the second control network are trained simultaneously and there is an error in the target state information output from the first control network, the target state information with the error is input to the second control network, and the animation processing model is trained inversely according to the joint action information output by the second control network, and as a result, the animation processing model is unstable and cannot efficiently process complex tasks. Therefore, in order to enable the animation processing model to handle complex tasks, the first control network to be trained and the second control network to be trained must be trained separately during training. Completing the training of the first control network to be trained is to obtain the first control network by training the first control network to be trained, which is connected to the second control network with fixed parameters.

In an embodiment of the present disclosure, the animation processing model is learned based on the AC algorithm framework, the first control network to be trained and the second control network to be trained in the animation processing model are trained separately, and thus the first control network to be trained The network and the second control network to be trained may each be set to include a pair of AC networks, that is, the first control network to be trained includes a first actor sub-network to be trained and a first critic sub-network to be trained; , the second control network to be trained includes a second training target actor sub-network and a second training target critic sub-network. Additionally, the structure of the first training target actor subnetwork may be set to be the same as the structure of the first training target critic subnetwork, and the structure of the second training target actor subnetwork may be set to be the same as the structure of the second training target critic subnetwork. It can be set as follows. Refer to FIG. 10 for the structure of the first training target actor sub-network and the first training target critic sub-network, and refer to FIG. 11 for the structure of the second training target actor sub-network and the second training target critic sub-network. You can; The difference lies only in the information input and the information output. After the training of the first control network to be trained and the second control network to be trained is completed, simply call the first actor sub-network and the second actor sub-network, using the first actor sub-network, input terrain features, Target state information a _H corresponding to the main joint is output according to the task information and state information, and joint action information a _L for all joints of the virtual character according to the target state information and state information using the second actor subnetwork. outputs.

An example is used as a human-shaped virtual character avoiding obstacles. During training of the second control network to be trained, training may be performed on flat ground using a set of animated segments. The animation segment set includes a plurality of animation segment samples covering different gestures of leg lifting and stepping of the virtual character in front of obstacles of different heights. The initial actions are close and have only one step. For example, an animation segment set has a total of 15 animation segment samples, and each animation segment sample is 0.5 seconds long. During training, the most appropriate animation segment sample may be selected from among a plurality of animation segment samples for training.

After obtaining a set of animation segments, mirroring can be performed for each animation segment sample, that is, the stepping of the character's right leg is converted to the stepping of the left leg, and the stepping of the character's left leg is converted to the stepping of the right leg. can be converted to stepping, and the initial gesture of each action segment and the landing a foot gesture are counted. Because the second control network uses the output of the first control network as the target task, when the second control network to be trained is trained, the initial gesture of the virtual character can be set in advance, and the target animation segment sample is applied to the initial gesture. Accordingly, the target task may be determined from the animation segment set, and the target task may be determined according to the target animation segment sample, so that the second control network to be trained learns according to the target animation segment sample. When the virtual character has completed one step and is ready to start the next step, the above-described operation can be repeated, and samples of target animation segments identical or similar to the initial gesture can be obtained to train a second control network to be trained. there is. Once the target animation segment sample is determined, the initial gesture may be compared with each animation segment sample in the animation segment set to obtain the similarity between the initial gesture and the gesture of the virtual character of each animation segment sample, and then the plurality of similarities may be ranked in descending order. Form a sequence by ranking, and finally use the animation segment sample corresponding to the highest similarity as the target animation segment sample; Alternatively, a preset number of similarities can be continuously obtained from the sequence, and any one of the animation segment samples corresponding to the similarity is used as the target animation segment sample, and the preset number can be set according to actual requirements, e.g. For example, the preset number may be 3 or 5. After the target animation segment sample is determined, state information samples corresponding to major joints are extracted from the target animation segment samples, the state information samples are used as target task information, and joint action information samples corresponding to all joints of the virtual character are simultaneously extracted. acquire; Then, training is performed by inputting the target task information into the second control network to be trained. If the joint action information output by the second control network to be trained is the same as or similar to the joint action information sample, it is the second control network to be trained. Indicates that training has been completed. A sample of joint action information is the rotation of both thighs of a kinematic character when landing on a foot. and rotation And the velocity direction of the root joint in the plane It can be, describes the rotation of the landing foot relative to the thigh joint, describes the rotation of the non-landing foot about to land/the rotation of the landing foot corresponding to the thigh joint, and the current target task information Rotation of both thighs and rotation and the velocity direction of the root joint in the plane It can be determined according to , and is input to the second control network to be trained to perform training.

To ensure the stability of the second control network, training may be performed according to a plurality of running paths of the character, and the maximum value of the character's running path is a numeric value such as 200 seconds (200s). It can be set to . The second control network to be trained based on reinforcement learning evaluates the action or state of the virtual character after the virtual character completes the running path, determines the state value, and performs actions according to the state value until the maximum state value is obtained. Adjust. By training, when different target task information is input, the second control network can perform a corresponding stepping action.

In an embodiment of the present disclosure, after the training of the second control network is completed, the first control network to be trained can be trained, and only one animation segment sample of the run is used for training, and the maximum length of each path is 200 seconds. It may be limited. When each path starts, the current moment's terrain feature sample, character state sample, and task information sample may be input, and the current moment's terrain feature sample, character state sample, and task information sample may be input using a first control network to be trained. Feature extraction is performed on the sample, and action information is output. Action information may be input to a second control network trained as a target task, and corresponding joint action information may be output by the second control network to control the character's actions. Similarly, after the virtual character completes the running path, the first control network to be trained based on reinforcement learning can determine the state value corresponding to the state of the virtual character according to the reward fed back by the environment, and the state value is When the preset value or maximum value is reached, this indicates that training of the first control network to be trained is complete.

In step S330, joint torque is determined according to joint action information, rendering is performed based on the joint torque to obtain gesture adjustment information corresponding to the virtual character, and animation segments are processed according to the gesture adjustment information.

In an embodiment of the present disclosure, after the joint action information output by the animation processing model is acquired, the joint torque may be determined according to the joint action information, and the joint torque may be additionally calculated using a physical engine to form a rigid body structure. It is applied to the joint corresponding to the structure and performs rendering to obtain gesture adjustment information corresponding to the virtual character, and the animation segment is processed according to the gesture adjustment information.

In embodiments of the present disclosure, in motion animation, character gestures are generally controlled using methods based on inverse kinematics. However, for physics-based character gesture control, when kinematic methods are used to control the character in real time, no real physical effects occur and interaction effects such as collisions are not perceived, and therefore, torque is generally used to control the character. Used to control actions. There are three main ways to control a physical character in real time: (1) Torque control: The model directly outputs the torque applied to each joint. This method is easy to implement. However, the control effect is weak, dynamic control is unstable, shaking is prone to occur, and the action is not natural enough. (2) Position control: the model provides target positions for each joint, and then the character is dynamically controlled to be in the corresponding position using a proportional-derivative controller (PD controller). Compared with torque control, position control is more stable, the model outputs the gesture of each joint, this method has relatively small distribution variance, small samples, and high model convergence speed. However, the existing PD control still has relatively large fluctuations. (3) Speed control: the model directly provides the target speed of each joint, the character is then dynamically controlled to the target speed using the PD control algorithm, and the effect of speed control and the model convergence speed are determined by the effect of position control and It is substantially consistent with the model convergence speed.

However, usually a position controller is used, which is equivalent to hierarchical control, the decision network obtains the current character state, outputs the target position for the next moment, and then the character uses the PD controller. It is dynamically controlled to be in the target gesture, and in actual engineering, the control cycle of PD is set to 100. This method has a positive effect on both model convergence speed and robustness. However, when using a typical PD controller, the shaking is relatively large and the gestures are not very standard.

An embodiment of the present disclosure provides stable PD control based on inverse kinematics to overcome the shortcomings of existing gesture control methods. The calculation formula for determining torque using an existing PD controller is as follows:

is the torque output, q is the current position of the virtual character's joints at the current moment, is the target position of the joint of the virtual character, is the speed of the joint at the current moment, k _p is the proportional coefficient, k _d is the differential gain coefficient, and n is the number of control cycles of PD control.

While controlling a physical character, the controller must quickly reduce the difference from the target gesture, so k _p should be set larger. In this case, stability problems with high proportional gain are likely to occur. Stable PD control can well solve this problem. Stable PD control is achieved in the following time period: using the location after It is obtained by calculating , which is equivalent to considering the initial state while comparing the difference from the target, thereby improving the stability of physical properties. In some embodiments, the current position and target position of the joint are determined according to joint action information; The current velocity and current acceleration of the joint are determined according to the current position, and the target velocity of the joint is determined according to the target position; According to the current velocity and the current acceleration, a first position and a first velocity corresponding to the joint are determined after the next control cycle; The joint torque is calculated according to the proportional coefficient, differential gain coefficient, current position, target position, target velocity, first position, and first velocity. Refer to equation (5) for the calculation formula.

is the torque output, k _p is the proportional coefficient, k _d is the differential gain coefficient, is the current location, is the first position of the joint at the current velocity after a certain period of time, is the target position of the joint, is the current velocity of the joint, is the first velocity of the joint at the current acceleration after a certain period of time, is the target speed of the joint, and n is the number of control cycles of the controller.

In an embodiment of the present disclosure, a plurality of torques corresponding to one joint action information may be determined using stable PD control based on inverse kinematics, and the plurality of torques may be additionally applied to the corresponding joints using a physical engine. The angular velocity and final gesture are calculated according to the rotation axis and anchor point of the joint, simulating the actual conditions of joint rotation, that is, gesture adjustment information corresponding to the virtual character at the current moment can be obtained, Gesture adjustment information may be an action sequence. Stable PD control based on inverse kinematics can improve calculation accuracy, reduce shaking, and improve the action effect of virtual characters.

In an embodiment of the present disclosure, the above-described solution is repeated in successive periods until the simulation of the image of the last frame of the animation segment is completed, so that at each moment the gesture coordination information corresponding to the virtual character, i.e. Gesture adjustment information corresponding to the virtual character may be obtained from the image frame of will be. The target action sequence may be determined according to gesture adjustment information corresponding to the virtual character at each moment. From the user's perspective, the animation effect presented by the target action sequence is more vivid than the effect of the original animation segment, the virtual character can avoid the newly set obstacles and complete the corresponding tasks, and the animation effect is more vivid. , the user experience is better.

Figures 14 (A) to (J) show action sequences of a virtual character controlled by an animation processing model and running on a flat surface. As shown in Figures 14 (A) to (J), the virtual character's leg lifting, stepping, foot landing, and arm shaking actions are more natural and vivid.

15A to 15E show action sequences of a human-shaped virtual character running on dense notched terrain. As shown in FIGS. 15A to 15E, two human-shaped virtual characters are included: a white human-shaped virtual character W and a black human-shaped virtual character B. The white human-shaped virtual character W is a human-shaped virtual character from the original animation segment, and the black human-shaped virtual character B is a human-shaped virtual character controlled by the animation processing model. 15A to 15E, the actions of the white human-shaped virtual character W and the black human-shaped virtual character B are the same, and at the notch C between the white human-shaped virtual character W and the black human-shaped virtual character B You can see that there is only a difference in stepping. The black human-shaped virtual character B, controlled by the animation processing model, can successfully complete the entire run of the dense notched terrain G, as shown in FIGS. 15A, 15B, 15D, and 15E.

16A to 16L show action sequences of a human-shaped virtual character running on a hybrid obstacle terrain. As shown in FIGS. 16A to 16L, the ground (G) of the hybrid obstacle terrain includes notches (C), protrusions (E), and steps (D). Similar to Figure 15, the diagram includes a white human-shaped virtual character W from the original animation segment and a black human-shaped virtual character B controlled by the animation processing model. Figures 16A to 16E show an action sequence of a human-shaped virtual character crossing a notch, Figures 16F to 16K show an action sequence of a human-shaped virtual character crossing a bump, and Figure 16L shows a human-shaped virtual character crossing a bump. represents an action sequence that crosses steps. It can be seen that the black human-shaped virtual character B can go over notches, bumps, and steps better, while the white human-shaped virtual character W has a poor running effect. For example, the feet of W, a white human-shaped virtual character, may be above notches or under protuberances or steps, which has an unrealistic animation effect.

The animation processing method according to an embodiment of the present disclosure can be applied to any game or animation design that requires physical animation. According to the animation processing method according to an embodiment of the present disclosure, an animation segment designed by an animator can be simulated. During simulation, obstacles and tasks can be set for the virtual character. Using the animation processing model, joint action information corresponding to the virtual character at the next moment is determined according to the terrain features and the task information and state information corresponding to the virtual character at the current moment. For example, the virtual user lands on his left foot and lifts his right foot at the current moment, the topographical feature is that there are bumps in the virtual user's movement path, and the task information is that the speed direction is forward, so the animation processing model is based on this information. Accordingly, the joint action information of the virtual character at the next moment is output so that the virtual character can successfully cross the bump after performing the action at multiple moments. Finally, the joint torque is determined according to the joint action information, and the joint torque is applied to the thigh and foot using a physical engine to perform rendering, thereby obtaining the action of the virtual character crossing the bump.

An animation processing method according to an embodiment of the present disclosure can be applied to any type of game animation. Augmented reality games are used as examples. Based on Figure 4, which shows a schematic diagram of the scene obtained after the integration of the game scene and the real scene, in the game animation the evil spirit (V) is a virtual character, the environment where the evil spirit is is a step (S) in the real scene, and behind the evil spirit is Electric scooters (M) are lined up. According to the animation processing method according to the embodiment of the present disclosure, the user can set a task so that the evil spirit (V) goes down the steps or bypasses the electric scooter (M). According to the status information and task information of the evil spirit (V) and the terrain characteristics of the graphical user interface, a vivid action sequence corresponding to the evil spirit can be obtained. In terms of visual effects, the demon (V) can jump from the current step to the next step or successfully bypass the electric scooter (M), and its feet may appear under the step (S) or overlap its body with the electric scooter (M). It doesn't happen. Therefore, the action becomes more vivid and self-adaptation to the environment becomes stronger.

According to the animation processing method according to an embodiment of the present disclosure, joint action information of the next instant adjacent to each instant is output according to terrain features and state information and task information of the virtual character at each instant using an animation processing model, and joint action By applying the joint torque determined according to the information to the corresponding joint using a physical engine and rendering it, a vivid action sequence can be obtained. The animation effects of animations created according to vivid action sequences are more natural and vivid than animations designed by animators. In addition, during the processing process, different terrains and tasks are added, and the interaction between the user and the virtual character is implemented in the game, so that the virtual character has self-adaptation, improving the virtual character's ability to recognize the terrain. By migrating the actions performed by a virtual character on flat ground to complex terrain, it is possible to improve the fun of the game, further improving the user experience and reducing the production cost of game animation.

The following describes device embodiments of the present disclosure that can be used to perform the animation processing method of the above-described embodiments of the present disclosure. For details not disclosed in the device embodiments of the present disclosure, refer to the above-described animation processing method of the present disclosure.

Figure 17 is a schematic block diagram of an animation processing device according to an embodiment of the present disclosure.

Referring to FIG. 17, the animation processing device 1700 according to an embodiment of the present disclosure includes an information acquisition module 1701, a model processing module 1702, and a gesture adjustment module 1703.

The information acquisition module 1701 is configured to acquire terrain features in the graphical user interface at the current moment, and obtain state information and task information corresponding to the virtual character in the animation segment at the current moment; The model processing module 1702 inputs terrain features, state information, and task information into the animation processing model, uses the animation processing model to perform feature extraction on the terrain features, state information, and task information, and performs feature extraction on the terrain features, state information, and task information to create a virtual character at the next moment. configured to obtain joint action information corresponding to; The gesture adjustment module 1703 determines the joint torque according to the joint action information, performs rendering based on the joint torque, obtains gesture adjustment information corresponding to the virtual character at the current moment, and segments animations according to the gesture adjustment information. It is configured to process.

In an embodiment of the present disclosure, the animation processing device 1700 further determines state information according to the gesture information of the virtual character of the initial moment of the animation segment when the current moment is the initial moment of the animation segment: If it is not the initial moment of the animation segment, the state information is determined according to the joint action information corresponding to the virtual character at the previous moment.

In an embodiment of the present disclosure, the animation processing device 1700 further acquires gesture adjustment information corresponding to the virtual character at a plurality of moments based on the animation segment; And it is configured to determine the target action sequence according to the gesture adjustment information at a plurality of moments.

In embodiments of the present disclosure, the terrain features are self-defined terrain features or actual terrain features; The state information includes the gesture, speed, and phase of each joint of the virtual character; The task information includes target speed direction or target point coordinates corresponding to the virtual character.

In an embodiment of the present disclosure, the animation processing model includes a first control network and a second control network. The model processing module 1702 inputs terrain features, state information, and task information into the first control network, and uses the first control network to perform feature extraction on the terrain features, state information, and tasks to generate features corresponding to major joints. a first feature extraction unit configured to obtain target state information; and using the target state information as target task information, inputting the state information and target task information into a second control network, and performing feature extraction on the state information and target task information using the second control network to obtain joint action information. and a second feature extraction unit configured to obtain.

In an embodiment of the present disclosure, the first control network includes a convolution unit, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The first feature extraction unit: performs feature extraction on the terrain features using a convolution unit to obtain first feature information corresponding to the terrain; Using the first fully connected layer, perform feature combination on the first feature information to obtain second feature information; Using the second fully connected layer, perform feature combination on the second feature information, state information, and task information to obtain third feature information; It is configured to obtain target state information by performing feature combination on the third feature information using the third fully connected layer.

In an embodiment of the present disclosure, the second control network includes a fourth fully connected layer and a fifth fully connected layer. The second feature extraction unit: performs feature combination on the state information and target task information using the fourth fully connected layer to obtain fourth feature information; And it is configured to obtain joint action information by performing feature combination on the fourth feature information using the fifth fully connected layer.

In an embodiment of the present disclosure, the gesture coordination module 1703: determines the current position and target position of the joint according to the joint action information; Determine the current velocity and current acceleration of the joint according to the current position, and determine the target velocity of the joint according to the target position; determine the first position and first velocity corresponding to the joint after the next control cycle according to the current velocity and current acceleration; It is configured to calculate the joint torque according to the proportional coefficient, differential gain coefficient, current position, target position, target speed, first position, and first speed.

In an embodiment of the present disclosure, the gesture coordination module 1703 is configured to input joint torque to the physical engine, apply the joint torque to the corresponding joint using the physical engine, and perform rendering to generate gesture coordination information. .

In an embodiment of the present disclosure, the animation processing device 1700: Before performing feature extraction on terrain features, state information, and task information using the animation processing model, train an animation processing model to be trained to create an animation processing model. It further includes a training module configured to obtain.

In an embodiment of the present disclosure, the animation processing model to be trained includes a first control network to be trained and a second control network to be trained; The training module is configured to obtain a terrain feature sample, a character state sample, and a task information sample, and train a first control network to be trained according to the terrain feature sample, the character state sample, and the task information sample, to obtain a first control network. 1 training unit; and a second training unit configured to train the second control network according to the state information sample corresponding to the main joint of the virtual character and the joint action information sample corresponding to all joints of the animation segment sample to obtain a second control network. , the first control network to be trained and the second control network to be trained are trained separately; And when the first control network to be trained is trained, the first control network to be trained is connected to the second control network with fixed parameters.

In an embodiment of the present disclosure, the second training unit acquires a plurality of animation segment samples, and determines a target animation segment sample from the plurality of animation segment samples according to the initial gesture of the virtual character; Obtain state information samples corresponding to major joints from the target animation segment samples, and use the state information samples as target task information; Obtain joint action information samples corresponding to all joints of the virtual character; And configured to train a second control network to be trained according to the target task information and joint action information samples.

In an embodiment of the present disclosure, the first control network to be trained includes a first actor sub-network to be trained and a first critical sub-network to be trained, and the second control network to be trained includes a second actor sub-network to be trained and a second to-be-trained actor sub-network. A critic subnetwork to be trained, wherein the structure of the first actor subnetwork to be trained is the same as the structure of the first critic subnetwork to be trained, and the structure of the second actor subnetwork to be trained is the second critic subnetwork to be trained. It is the same as the structure of

Figure 18 shows a schematic structural diagram of a computer system of an electronic device adapted to implement embodiments of the present disclosure.

The computer system 1800 of the electronic device shown in FIG. 18 is only an example and does not limit the function and scope of use of the embodiments of the present disclosure.

As shown in FIG. 18, the computer system 1800 includes a central processing unit (CPU) 1801, which stores a program or program stored in a read-only memory (ROM) 1802. The animation processing method of the above-described embodiment is implemented by performing various appropriate actions or processes according to the program loaded into the random access memory (RAM) 1803 from the storage part 1808. RAM 1803 additionally stores various programs and data necessary for system operation. CPU 1801, ROM 1802, and RAM 1803 are connected to each other through a bus 1804. Input/output (I/O) interface 1805 is also coupled to bus 1804.

An input part 1806 including a keyboard, a mouse, etc., an output part 1807 including a cathode ray tube (CRT), a liquid crystal display (LCD), a speaker, etc., a hard disk, etc. Components including a storage unit 1808 including a and a communication unit 1809 including a network interface card such as a local area network (LAN) card or modem are connected to the I/O interface 1805. . The communication unit 1809 performs communication processing using a network such as the Internet. Drive 1810 is also connected to I/O interface 1805 as needed. Since a removable medium 1811 such as a magnetic disk, optical disk, magneto-optical disk, or semiconductor memory is installed in the drive 1810 as needed, the computer program read from the removable medium can be stored in the storage unit as needed. Installed in 1808.

In particular, according to an embodiment of the present disclosure, processes described later with reference to the flowchart may be implemented as a computer software program. For example, this embodiment of the present disclosure includes a computer program product, the computer program product includes a computer program carried on a computer-readable medium, and the computer program is used to perform the method shown in the flowchart. Contains program code. In this embodiment, by using the communication unit 1809, a computer program can be downloaded and installed from a network and/or installed from a removable medium 1811. When a computer program is executed by the CPU 1801, various functions defined in the system of the present application are executed.

In embodiments of the present disclosure, the computer-readable medium may be a computer-readable signal medium, a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. Computer-readable storage media include electrical connections with one or more wires, portable computer magnetic disks, hard disks, random access memory (RAM), read-only memory (ROM), and erasable programs. erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any of these. It may include an appropriate combination of, but is not limited to this. In the present disclosure, the computer-readable storage medium may be any tangible medium that includes or stores a program, and the program may be used by or in combination with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium may include a data signal in baseband or propagating as part of a carrier wave, where the data signal carries computer-readable program code. Data signals propagated in this manner may take multiple forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium may further include any computer-readable medium in addition to the computer-readable storage medium. A computer-readable medium can transmit, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. Program code included in a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless medium, wired, etc., or any suitable combination thereof.

The flow diagrams and block diagrams in the accompanying drawings illustrate possible system architectures, functions, and operations that may be implemented by systems, methods, and computer program products in accordance with various embodiments of the present disclosure. In this regard, each box in a flowchart or block diagram may represent a module, program segment, or portion of code. A module, program segment, or portion of code contains one or more executable instructions used to implement specified logical functions. In some alternative implementations, the functions annotated in the box may occur in a different order than those annotated in the accompanying drawings. For example, two boxes that are actually shown back-to-back may essentially be performed in parallel, and sometimes two boxes may be performed in reverse order. This is determined by the relevant function. Each box in the block diagram and/or flow diagram and combination of boxes in the block diagram and/or flow diagram may be implemented using a dedicated hardware-based system configured to perform the specified function or operation, or using a combination of dedicated hardware and computer instructions. It can be implemented.

Related units described in embodiments of the present disclosure may be implemented in software or hardware, and the described units may be set in a processor. The name of the unit does not constitute a restriction to the unit in any particular case.

According to another aspect, the present disclosure further provides a computer-readable medium. The computer-readable medium may be included in the animation processing device described in the above-described embodiment, or may exist alone and not be disposed in the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to implement the methods described in the foregoing embodiments.

Although the foregoing detailed description has discussed multiple modules or units of the device configured to perform actions, such distinction is not required. In fact, according to an implementation example of the present disclosure, the features and functions of two or more modules or units described above may be implemented with one module or unit. Conversely, the features and functions of one module or unit described above may be further divided and implemented as a plurality of modules or units.

Following the foregoing description of the implementation, one skilled in the art will readily understand that the example implementations described herein may be implemented using software, or may be implemented by combining software with the necessary hardware. Accordingly, the technical solution according to an embodiment of the present disclosure may be implemented in the form of a software product. The software product is stored on a non-volatile storage medium (which may be a CD-ROM, USB flash drive, removable hard disk, etc.) or a network, and is stored on a computing device (which may be a personal computer, server, touch terminal, network device, etc.) disclosed herein. Includes several commands to perform the method according to the embodiment.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure herein. This disclosure is intended to cover any modification, use, or adaptation of this disclosure. Such modifications, uses or adaptive changes follow the general principles of the present disclosure and include in the technique common sense or technical means not disclosed in the present disclosure.

It should be understood that the present disclosure is not limited to the exact structure described above and shown in the accompanying drawings, and that various modifications and changes may be made without departing from the scope of the present disclosure. The scope of the present disclosure is limited only by the appended claims.

Claims (16)

An animation processing method applicable to electronic devices, comprising:
Obtaining a terrain feature from a graphical user interface of a current moment, and obtaining state information and task information corresponding to a virtual character in an animation segment of the current moment;
Input the terrain features, the state information, and the task information into an animation processing model, perform feature extraction on the terrain features, the state information, and the task information using the animation processing model, and perform feature extraction on the terrain features, the state information, and the task information, and Obtaining joint action information corresponding to the character;
determining joint torque according to the joint action information; and
Obtaining gesture adjustment information corresponding to the virtual character at the current moment based on the joint torque, and processing the animation segment according to the gesture adjustment information,
The animation processing model includes a first control network and a second control network;
Input the terrain features, the state information, and the task information into an animation processing model, perform feature extraction on the terrain features, the state information, and the task information using the animation processing model, and perform feature extraction on the terrain features, the state information, and the task information, and The step of acquiring joint action information corresponding to the character is,
Input the terrain features, the state information, and the task information into the first control network, perform feature extraction on the terrain features, the state information, and the task information using the first control network, and perform main ( key) Obtaining target state information corresponding to a joint - the main joint corresponds to the terrain feature and the state information and task information of the virtual character -
determining the target state information as target task information; and
Inputting the state information and the target task information into the second control network, and performing feature extraction on the state information and the target task information using the second control network to obtain the joint action information.
An animation processing method containing . According to paragraph 1,
If the current moment is the initial moment of the animation segment, determining the state information according to gesture information of the virtual character at the initial moment of the animation segment; and
If the current moment is not the initial moment of the animation segment, determining the state information according to joint action information corresponding to the virtual character at a previous point in time.
An animation processing method further comprising: According to paragraph 2,
Obtaining gesture adjustment information corresponding to the virtual character at a plurality of moments based on the animation segment; and
Determining a target action sequence according to the gesture adjustment information at the plurality of moments
An animation processing method further comprising: According to paragraph 1,
The terrain features are self-defined terrain features or actual terrain features;
The state information includes the gesture, speed, and phase of each joint of the virtual character; and
The animation processing method wherein the task information includes a target speed direction or target point coordinates corresponding to the virtual character. delete According to paragraph 1,
The first control network includes a convolution unit, a first fully connected layer, a second fully connected layer, and a third fully connected layer; and
The step of performing feature extraction on the terrain feature, the state information, and the task information using the first control network to obtain target state information corresponding to a major joint,
performing feature extraction on the terrain features using the convolution unit to obtain first feature information corresponding to the terrain;
Obtaining second feature information by performing feature combination on the first feature information using the first fully connected layer;
Obtaining third feature information by performing feature combination on the second feature information, the state information, and the task information using the second fully connected layer; and
Obtaining the target state information by performing feature combination on the third feature information using the third fully connected layer.
An animation processing method, including. According to paragraph 1,
the second control network includes a fourth fully connected layer and a fifth fully connected layer; and
Obtaining the joint action information by performing feature extraction on the state information and the target task information using the second control network,
Obtaining fourth feature information by performing feature combination on the state information and the target task information using the fourth fully connected layer; and
Obtaining the joint action information by performing feature combination on the fourth feature information using the fifth fully connected layer
An animation processing method including. According to paragraph 1,
The step of determining joint torque according to the joint action information is:
determining the current position and target position of the joint according to the joint action information;
determining the current velocity and current acceleration of the joint according to the current position, and determining the target velocity of the joint according to the target position;
determining a first position and a first velocity corresponding to the joint after a next control period according to the current velocity and the current acceleration; and
calculating the joint torque according to a proportional coefficient, a differential gain coefficient, the current position, the target position, the target velocity, the first position, and the first velocity.
An animation processing method including. According to paragraph 1,
Obtaining gesture adjustment information corresponding to the virtual character at the current moment based on the joint torque,
Inputting the joint torque into a physics engine, applying the joint torque to a corresponding joint using the physics engine, and performing rendering to generate the gesture adjustment information.
An animation processing method including. According to paragraph 1,
Before performing feature extraction on the terrain features, the state information, and the task information using the animation processing model,
The animation processing method is,
Obtaining the animation processing model by training a to-be-trained animation processing model.
An animation processing method further comprising: According to clause 10,
The animation processing model to be trained includes a first control network to be trained and a second control network to be trained; and
The step of training the training target animation processing model to obtain the animation processing model,
acquiring a terrain feature sample, a character state sample, and a task information sample;
training the first control network to be trained according to the terrain feature sample, the character state sample, and the task information sample to obtain the first control network; and
Obtaining the second control network by training the second control network according to state information samples corresponding to main joints of the virtual character and joint action information samples corresponding to all joints of the animation segment samples.
Including,
the first control network to be trained and the second control network to be trained are trained separately; When the first control network to be trained is trained, the first control network to be trained is connected to the second control network with fixed parameters. According to clause 11,
Obtaining the second control network by training the second control network according to the state information sample corresponding to the main joint of the virtual character and the joint action information sample corresponding to all joints of the animation segment sample,
Obtaining a plurality of animation segment samples;
determining a target animation segment sample from the plurality of animation segment samples according to an initial gesture of the virtual character;
Obtaining a state information sample corresponding to the main joint from the target animation segment sample and using the state information sample as target task information;
Obtaining the joint action information samples corresponding to all joints of the virtual character; and
Training the second control network to be trained according to the target task information and the joint action information sample.
An animation processing method including. According to clause 11,
The first control network to be trained includes a first training target actor sub-network and a first training target critical sub-network; and
The second control network to be trained includes a second actor sub-network to be trained and a second critic sub-network to be trained,
An animation processing method wherein the structure of the first actor sub-network to be trained is the same as the first critic sub-network to be trained, and the structure of the second actor sub-network to be trained is the same as the structure of the critique sub-network to be trained. . An animation processing device, comprising:
an information acquisition module configured to obtain terrain features in a graphical user interface at the current moment and obtain state information and task information corresponding to a virtual character in an animation segment at the current moment;
Input the terrain features, the state information, and the task information into an animation processing model, perform feature extraction on the terrain features, the state information, and the task information using the animation processing model, and perform feature extraction on the terrain features, the state information, and the task information, and a model processing module configured to acquire joint action information corresponding to the character; and
Gesture adjustment configured to determine a joint torque according to the joint action information, obtain gesture adjustment information corresponding to the virtual character at the current moment based on the joint torque, and process the animation segment according to the gesture adjustment information. Contains modules,
The animation processing model includes a first control network and a second control network;
The model processing module is,
Input the terrain features, the state information, and the task information into the first control network, perform feature extraction on the terrain features, the state information, and the task information using the first control network, and perform main ( key) a first feature extraction unit configured to obtain target state information corresponding to a joint, wherein the main joint corresponds to the terrain feature and state information and task information of the virtual character; and
Determine the target state information as target task information, input the state information and the target task information into the second control network, and extract features for the state information and the target task information using the second control network. A second feature extraction unit configured to obtain the joint action information by performing
An animation processing unit comprising: As an electronic device,
One or more processors; and
A storage device configured to store one or more programs
Includes,
The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the animation processing method according to any one of claims 1 to 4 and 6 to 13, an electronic device. . A computer-readable storage medium that stores a computer program,
A computer-readable storage medium, wherein the computer program, when executed by a processor, implements the animation processing method according to any one of claims 1 to 4 and 6 to 13.